Developing intelligent decision making systems in the real world requires planning algorithms which are able to deal with sources of uncertainty and constraints. An example can be found in smart distribution grids, in which planning can be used to decide when electric vehicles charge their batteries, such that the capacity limits of lines are respected at all times. In this particular example there can be uncertainty in the arrival time and charging demand of vehicles, and constraints follow directly from the capacity limits of the distribution grid to which vehicles are connected. Existing algorithms for planning under uncertainty subject to constraints are currently not suitable for these types of applications, and therefore this dissertation aims improve the applicability of these algorithms by advancing the state of the art in constrained multi-agent planning under uncertainty. The dissertation presents new algorithmic techniques for exact POMDP planning, finite-horizon POMDPs and POMDPs with constraints. Additionally, the dissertation shows how models for constrained planning can be used in smart distribution grids.

The current architecture of the power grid is outdated and will not provide the means to deal with the decentralization of energy sources. The smart grid is a newly envisioned architecture for the power grid that should solve the weaknesses in the current grid. One application that is part of the smart grid vision is the microgrid: a coalition of prosumers that is able to operate either in parallel to the power grid or isolated from it.

In this work we investigate how these microgrids should be formed in order to maximize their usability while at the same time preserving the freedom of choice for participants. We apply a game theoretic perspective to simulate the choices of individual users. This results in a problem that combines the notion of core stability and maximization of social welfare. For this problem we provide a complexity proof, followed by a relaxation of core stability, $k$-stability, by limiting the available knowledge for participants. Finally we produce our own distributed algorithm which can act as a heuristic.

In recent years,we have seen amajor shift in computing systems: data volumes are growing very fast, but hardware capabilities to store, process, and transfer the massive data are not speeding up at the same rate. Today, data are generated from a variety of sources, such as social networking websites, business transactions, banking sectors, etc. These data are valuable and contain lots of vital information if they are analyzed efficiently. The processing capabilities of single machines, however, are not sufficient enough, which
makes it harder to use them for data analysis. As a result, most web companies, but also the traditional business organizations, research labs, and universities, are scaling out their major computational frameworks to clusters of thousands of machines. To find the hidden and interesting insights from the data, in addition to simple queries, also complex machine learning algorithms and graphs processing are becoming a common choice in many areas. Nowadays, the problem to collect, store and analyze these data is called the Big Data problem.